tools/perf/util/machine.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <dirent.h>
   3 #include <errno.h>
   4 #include <inttypes.h>
   5 #include <regex.h>
   6 #include <stdlib.h>
   7 #include "callchain.h"
   8 #include "debug.h"
   9 #include "dso.h"
  10 #include "env.h"
  11 #include "event.h"
  12 #include "evsel.h"
  13 #include "hist.h"
  14 #include "machine.h"
  15 #include "map.h"
  16 #include "map_symbol.h"
  17 #include "branch.h"
  18 #include "mem-events.h"
  19 #include "srcline.h"
  20 #include "symbol.h"
  21 #include "sort.h"
  22 #include "strlist.h"
  23 #include "target.h"
  24 #include "thread.h"
  25 #include "util.h"
  26 #include "vdso.h"
  27 #include <stdbool.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <unistd.h>
  31 #include "unwind.h"
  32 #include "linux/hash.h"
  33 #include "asm/bug.h"
  34 #include "bpf-event.h"
  35 #include <internal/lib.h> // page_size
  36
  37 #include <linux/ctype.h>
  38 #include <symbol/kallsyms.h>
  39 #include <linux/mman.h>
  40 #include <linux/string.h>
  41 #include <linux/zalloc.h>
  42
  43 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  44
  45 static void dsos__init(struct dsos *dsos)
  46 {
  47         INIT_LIST_HEAD(&dsos->head);
  48         dsos->root = RB_ROOT;
  49         init_rwsem(&dsos->lock);
  50 }
  51
  52 static void machine__threads_init(struct machine *machine)
  53 {
  54         int i;
  55
  56         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  57                 struct threads *threads = &machine->threads[i];
  58                 threads->entries = RB_ROOT_CACHED;
  59                 init_rwsem(&threads->lock);
  60                 threads->nr = 0;
  61                 INIT_LIST_HEAD(&threads->dead);
  62                 threads->last_match = NULL;
  63         }
  64 }
  65
  66 static int machine__set_mmap_name(struct machine *machine)
  67 {
  68         if (machine__is_host(machine))
  69                 machine->mmap_name = strdup("[kernel.kallsyms]");
  70         else if (machine__is_default_guest(machine))
  71                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  72         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  73                           machine->pid) < 0)
  74                 machine->mmap_name = NULL;
  75
  76         return machine->mmap_name ? 0 : -ENOMEM;
  77 }
  78
  79 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  80 {
  81         int err = -ENOMEM;
  82
  83         memset(machine, 0, sizeof(*machine));
  84         map_groups__init(&machine->kmaps, machine);
  85         RB_CLEAR_NODE(&machine->rb_node);
  86         dsos__init(&machine->dsos);
  87
  88         machine__threads_init(machine);
  89
  90         machine->vdso_info = NULL;
  91         machine->env = NULL;
  92
  93         machine->pid = pid;
  94
  95         machine->id_hdr_size = 0;
  96         machine->kptr_restrict_warned = false;
  97         machine->comm_exec = false;
  98         machine->kernel_start = 0;
  99         machine->vmlinux_map = NULL;
 100
 101         machine->root_dir = strdup(root_dir);
 102         if (machine->root_dir == NULL)
 103                 return -ENOMEM;
 104
 105         if (machine__set_mmap_name(machine))
 106                 goto out;
 107
 108         if (pid != HOST_KERNEL_ID) {
 109                 struct thread *thread = machine__findnew_thread(machine, -1,
 110                                                                 pid);
 111                 char comm[64];
 112
 113                 if (thread == NULL)
 114                         goto out;
 115
 116                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 117                 thread__set_comm(thread, comm, 0);
 118                 thread__put(thread);
 119         }
 120
 121         machine->current_tid = NULL;
 122         err = 0;
 123
 124 out:
 125         if (err) {
 126                 zfree(&machine->root_dir);
 127                 zfree(&machine->mmap_name);
 128         }
 129         return 0;
 130 }
 131
 132 struct machine *machine__new_host(void)
 133 {
 134         struct machine *machine = malloc(sizeof(*machine));
 135
 136         if (machine != NULL) {
 137                 machine__init(machine, "", HOST_KERNEL_ID);
 138
 139                 if (machine__create_kernel_maps(machine) < 0)
 140                         goto out_delete;
 141         }
 142
 143         return machine;
 144 out_delete:
 145         free(machine);
 146         return NULL;
 147 }
 148
 149 struct machine *machine__new_kallsyms(void)
 150 {
 151         struct machine *machine = machine__new_host();
 152         /*
 153          * FIXME:
 154          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 155          *    ask for not using the kcore parsing code, once this one is fixed
 156          *    to create a map per module.
 157          */
 158         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 159                 machine__delete(machine);
 160                 machine = NULL;
 161         }
 162
 163         return machine;
 164 }
 165
 166 static void dsos__purge(struct dsos *dsos)
 167 {
 168         struct dso *pos, *n;
 169
 170         down_write(&dsos->lock);
 171
 172         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 173                 RB_CLEAR_NODE(&pos->rb_node);
 174                 pos->root = NULL;
 175                 list_del_init(&pos->node);
 176                 dso__put(pos);
 177         }
 178
 179         up_write(&dsos->lock);
 180 }
 181
 182 static void dsos__exit(struct dsos *dsos)
 183 {
 184         dsos__purge(dsos);
 185         exit_rwsem(&dsos->lock);
 186 }
 187
 188 void machine__delete_threads(struct machine *machine)
 189 {
 190         struct rb_node *nd;
 191         int i;
 192
 193         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 194                 struct threads *threads = &machine->threads[i];
 195                 down_write(&threads->lock);
 196                 nd = rb_first_cached(&threads->entries);
 197                 while (nd) {
 198                         struct thread *t = rb_entry(nd, struct thread, rb_node);
 199
 200                         nd = rb_next(nd);
 201                         __machine__remove_thread(machine, t, false);
 202                 }
 203                 up_write(&threads->lock);
 204         }
 205 }
 206
 207 void machine__exit(struct machine *machine)
 208 {
 209         int i;
 210
 211         if (machine == NULL)
 212                 return;
 213
 214         machine__destroy_kernel_maps(machine);
 215         map_groups__exit(&machine->kmaps);
 216         dsos__exit(&machine->dsos);
 217         machine__exit_vdso(machine);
 218         zfree(&machine->root_dir);
 219         zfree(&machine->mmap_name);
 220         zfree(&machine->current_tid);
 221
 222         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 223                 struct threads *threads = &machine->threads[i];
 224                 struct thread *thread, *n;
 225                 /*
 226                  * Forget about the dead, at this point whatever threads were
 227                  * left in the dead lists better have a reference count taken
 228                  * by who is using them, and then, when they drop those references
 229                  * and it finally hits zero, thread__put() will check and see that
 230                  * its not in the dead threads list and will not try to remove it
 231                  * from there, just calling thread__delete() straight away.
 232                  */
 233                 list_for_each_entry_safe(thread, n, &threads->dead, node)
 234                         list_del_init(&thread->node);
 235
 236                 exit_rwsem(&threads->lock);
 237         }
 238 }
 239
 240 void machine__delete(struct machine *machine)
 241 {
 242         if (machine) {
 243                 machine__exit(machine);
 244                 free(machine);
 245         }
 246 }
 247
 248 void machines__init(struct machines *machines)
 249 {
 250         machine__init(&machines->host, "", HOST_KERNEL_ID);
 251         machines->guests = RB_ROOT_CACHED;
 252 }
 253
 254 void machines__exit(struct machines *machines)
 255 {
 256         machine__exit(&machines->host);
 257         /* XXX exit guest */
 258 }
 259
 260 struct machine *machines__add(struct machines *machines, pid_t pid,
 261                               const char *root_dir)
 262 {
 263         struct rb_node **p = &machines->guests.rb_root.rb_node;
 264         struct rb_node *parent = NULL;
 265         struct machine *pos, *machine = malloc(sizeof(*machine));
 266         bool leftmost = true;
 267
 268         if (machine == NULL)
 269                 return NULL;
 270
 271         if (machine__init(machine, root_dir, pid) != 0) {
 272                 free(machine);
 273                 return NULL;
 274         }
 275
 276         while (*p != NULL) {
 277                 parent = *p;
 278                 pos = rb_entry(parent, struct machine, rb_node);
 279                 if (pid < pos->pid)
 280                         p = &(*p)->rb_left;
 281                 else {
 282                         p = &(*p)->rb_right;
 283                         leftmost = false;
 284                 }
 285         }
 286
 287         rb_link_node(&machine->rb_node, parent, p);
 288         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 289
 290         return machine;
 291 }
 292
 293 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 294 {
 295         struct rb_node *nd;
 296
 297         machines->host.comm_exec = comm_exec;
 298
 299         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 300                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 301
 302                 machine->comm_exec = comm_exec;
 303         }
 304 }
 305
 306 struct machine *machines__find(struct machines *machines, pid_t pid)
 307 {
 308         struct rb_node **p = &machines->guests.rb_root.rb_node;
 309         struct rb_node *parent = NULL;
 310         struct machine *machine;
 311         struct machine *default_machine = NULL;
 312
 313         if (pid == HOST_KERNEL_ID)
 314                 return &machines->host;
 315
 316         while (*p != NULL) {
 317                 parent = *p;
 318                 machine = rb_entry(parent, struct machine, rb_node);
 319                 if (pid < machine->pid)
 320                         p = &(*p)->rb_left;
 321                 else if (pid > machine->pid)
 322                         p = &(*p)->rb_right;
 323                 else
 324                         return machine;
 325                 if (!machine->pid)
 326                         default_machine = machine;
 327         }
 328
 329         return default_machine;
 330 }
 331
 332 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 333 {
 334         char path[PATH_MAX];
 335         const char *root_dir = "";
 336         struct machine *machine = machines__find(machines, pid);
 337
 338         if (machine && (machine->pid == pid))
 339                 goto out;
 340
 341         if ((pid != HOST_KERNEL_ID) &&
 342             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 343             (symbol_conf.guestmount)) {
 344                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 345                 if (access(path, R_OK)) {
 346                         static struct strlist *seen;
 347
 348                         if (!seen)
 349                                 seen = strlist__new(NULL, NULL);
 350
 351                         if (!strlist__has_entry(seen, path)) {
 352                                 pr_err("Can't access file %s\n", path);
 353                                 strlist__add(seen, path);
 354                         }
 355                         machine = NULL;
 356                         goto out;
 357                 }
 358                 root_dir = path;
 359         }
 360
 361         machine = machines__add(machines, pid, root_dir);
 362 out:
 363         return machine;
 364 }
 365
 366 void machines__process_guests(struct machines *machines,
 367                               machine__process_t process, void *data)
 368 {
 369         struct rb_node *nd;
 370
 371         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 372                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 373                 process(pos, data);
 374         }
 375 }
 376
 377 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 378 {
 379         struct rb_node *node;
 380         struct machine *machine;
 381
 382         machines->host.id_hdr_size = id_hdr_size;
 383
 384         for (node = rb_first_cached(&machines->guests); node;
 385              node = rb_next(node)) {
 386                 machine = rb_entry(node, struct machine, rb_node);
 387                 machine->id_hdr_size = id_hdr_size;
 388         }
 389
 390         return;
 391 }
 392
 393 static void machine__update_thread_pid(struct machine *machine,
 394                                        struct thread *th, pid_t pid)
 395 {
 396         struct thread *leader;
 397
 398         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 399                 return;
 400
 401         th->pid_ = pid;
 402
 403         if (th->pid_ == th->tid)
 404                 return;
 405
 406         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 407         if (!leader)
 408                 goto out_err;
 409
 410         if (!leader->mg)
 411                 leader->mg = map_groups__new(machine);
 412
 413         if (!leader->mg)
 414                 goto out_err;
 415
 416         if (th->mg == leader->mg)
 417                 return;
 418
 419         if (th->mg) {
 420                 /*
 421                  * Maps are created from MMAP events which provide the pid and
 422                  * tid.  Consequently there never should be any maps on a thread
 423                  * with an unknown pid.  Just print an error if there are.
 424                  */
 425                 if (!map_groups__empty(th->mg))
 426                         pr_err("Discarding thread maps for %d:%d\n",
 427                                th->pid_, th->tid);
 428                 map_groups__put(th->mg);
 429         }
 430
 431         th->mg = map_groups__get(leader->mg);
 432 out_put:
 433         thread__put(leader);
 434         return;
 435 out_err:
 436         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 437         goto out_put;
 438 }
 439
 440 /*
 441  * Front-end cache - TID lookups come in blocks,
 442  * so most of the time we dont have to look up
 443  * the full rbtree:
 444  */
 445 static struct thread*
 446 __threads__get_last_match(struct threads *threads, struct machine *machine,
 447                           int pid, int tid)
 448 {
 449         struct thread *th;
 450
 451         th = threads->last_match;
 452         if (th != NULL) {
 453                 if (th->tid == tid) {
 454                         machine__update_thread_pid(machine, th, pid);
 455                         return thread__get(th);
 456                 }
 457
 458                 threads->last_match = NULL;
 459         }
 460
 461         return NULL;
 462 }
 463
 464 static struct thread*
 465 threads__get_last_match(struct threads *threads, struct machine *machine,
 466                         int pid, int tid)
 467 {
 468         struct thread *th = NULL;
 469
 470         if (perf_singlethreaded)
 471                 th = __threads__get_last_match(threads, machine, pid, tid);
 472
 473         return th;
 474 }
 475
 476 static void
 477 __threads__set_last_match(struct threads *threads, struct thread *th)
 478 {
 479         threads->last_match = th;
 480 }
 481
 482 static void
 483 threads__set_last_match(struct threads *threads, struct thread *th)
 484 {
 485         if (perf_singlethreaded)
 486                 __threads__set_last_match(threads, th);
 487 }
 488
 489 /*
 490  * Caller must eventually drop thread->refcnt returned with a successful
 491  * lookup/new thread inserted.
 492  */
 493 static struct thread *____machine__findnew_thread(struct machine *machine,
 494                                                   struct threads *threads,
 495                                                   pid_t pid, pid_t tid,
 496                                                   bool create)
 497 {
 498         struct rb_node **p = &threads->entries.rb_root.rb_node;
 499         struct rb_node *parent = NULL;
 500         struct thread *th;
 501         bool leftmost = true;
 502
 503         th = threads__get_last_match(threads, machine, pid, tid);
 504         if (th)
 505                 return th;
 506
 507         while (*p != NULL) {
 508                 parent = *p;
 509                 th = rb_entry(parent, struct thread, rb_node);
 510
 511                 if (th->tid == tid) {
 512                         threads__set_last_match(threads, th);
 513                         machine__update_thread_pid(machine, th, pid);
 514                         return thread__get(th);
 515                 }
 516
 517                 if (tid < th->tid)
 518                         p = &(*p)->rb_left;
 519                 else {
 520                         p = &(*p)->rb_right;
 521                         leftmost = false;
 522                 }
 523         }
 524
 525         if (!create)
 526                 return NULL;
 527
 528         th = thread__new(pid, tid);
 529         if (th != NULL) {
 530                 rb_link_node(&th->rb_node, parent, p);
 531                 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
 532
 533                 /*
 534                  * We have to initialize map_groups separately
 535                  * after rb tree is updated.
 536                  *
 537                  * The reason is that we call machine__findnew_thread
 538                  * within thread__init_map_groups to find the thread
 539                  * leader and that would screwed the rb tree.
 540                  */
 541                 if (thread__init_map_groups(th, machine)) {
 542                         rb_erase_cached(&th->rb_node, &threads->entries);
 543                         RB_CLEAR_NODE(&th->rb_node);
 544                         thread__put(th);
 545                         return NULL;
 546                 }
 547                 /*
 548                  * It is now in the rbtree, get a ref
 549                  */
 550                 thread__get(th);
 551                 threads__set_last_match(threads, th);
 552                 ++threads->nr;
 553         }
 554
 555         return th;
 556 }
 557
 558 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 559 {
 560         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 561 }
 562
 563 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 564                                        pid_t tid)
 565 {
 566         struct threads *threads = machine__threads(machine, tid);
 567         struct thread *th;
 568
 569         down_write(&threads->lock);
 570         th = __machine__findnew_thread(machine, pid, tid);
 571         up_write(&threads->lock);
 572         return th;
 573 }
 574
 575 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 576                                     pid_t tid)
 577 {
 578         struct threads *threads = machine__threads(machine, tid);
 579         struct thread *th;
 580
 581         down_read(&threads->lock);
 582         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 583         up_read(&threads->lock);
 584         return th;
 585 }
 586
 587 struct comm *machine__thread_exec_comm(struct machine *machine,
 588                                        struct thread *thread)
 589 {
 590         if (machine->comm_exec)
 591                 return thread__exec_comm(thread);
 592         else
 593                 return thread__comm(thread);
 594 }
 595
 596 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 597                                 struct perf_sample *sample)
 598 {
 599         struct thread *thread = machine__findnew_thread(machine,
 600                                                         event->comm.pid,
 601                                                         event->comm.tid);
 602         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 603         int err = 0;
 604
 605         if (exec)
 606                 machine->comm_exec = true;
 607
 608         if (dump_trace)
 609                 perf_event__fprintf_comm(event, stdout);
 610
 611         if (thread == NULL ||
 612             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 613                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 614                 err = -1;
 615         }
 616
 617         thread__put(thread);
 618
 619         return err;
 620 }
 621
 622 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 623                                       union perf_event *event,
 624                                       struct perf_sample *sample __maybe_unused)
 625 {
 626         struct thread *thread = machine__findnew_thread(machine,
 627                                                         event->namespaces.pid,
 628                                                         event->namespaces.tid);
 629         int err = 0;
 630
 631         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 632                   "\nWARNING: kernel seems to support more namespaces than perf"
 633                   " tool.\nTry updating the perf tool..\n\n");
 634
 635         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 636                   "\nWARNING: perf tool seems to support more namespaces than"
 637                   " the kernel.\nTry updating the kernel..\n\n");
 638
 639         if (dump_trace)
 640                 perf_event__fprintf_namespaces(event, stdout);
 641
 642         if (thread == NULL ||
 643             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 644                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 645                 err = -1;
 646         }
 647
 648         thread__put(thread);
 649
 650         return err;
 651 }
 652
 653 int machine__process_lost_event(struct machine *machine __maybe_unused,
 654                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 655 {
 656         dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
 657                     event->lost.id, event->lost.lost);
 658         return 0;
 659 }
 660
 661 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 662                                         union perf_event *event, struct perf_sample *sample)
 663 {
 664         dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
 665                     sample->id, event->lost_samples.lost);
 666         return 0;
 667 }
 668
 669 static struct dso *machine__findnew_module_dso(struct machine *machine,
 670                                                struct kmod_path *m,
 671                                                const char *filename)
 672 {
 673         struct dso *dso;
 674
 675         down_write(&machine->dsos.lock);
 676
 677         dso = __dsos__find(&machine->dsos, m->name, true);
 678         if (!dso) {
 679                 dso = __dsos__addnew(&machine->dsos, m->name);
 680                 if (dso == NULL)
 681                         goto out_unlock;
 682
 683                 dso__set_module_info(dso, m, machine);
 684                 dso__set_long_name(dso, strdup(filename), true);
 685         }
 686
 687         dso__get(dso);
 688 out_unlock:
 689         up_write(&machine->dsos.lock);
 690         return dso;
 691 }
 692
 693 int machine__process_aux_event(struct machine *machine __maybe_unused,
 694                                union perf_event *event)
 695 {
 696         if (dump_trace)
 697                 perf_event__fprintf_aux(event, stdout);
 698         return 0;
 699 }
 700
 701 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 702                                         union perf_event *event)
 703 {
 704         if (dump_trace)
 705                 perf_event__fprintf_itrace_start(event, stdout);
 706         return 0;
 707 }
 708
 709 int machine__process_switch_event(struct machine *machine __maybe_unused,
 710                                   union perf_event *event)
 711 {
 712         if (dump_trace)
 713                 perf_event__fprintf_switch(event, stdout);
 714         return 0;
 715 }
 716
 717 static int machine__process_ksymbol_register(struct machine *machine,
 718                                              union perf_event *event,
 719                                              struct perf_sample *sample __maybe_unused)
 720 {
 721         struct symbol *sym;
 722         struct map *map;
 723
 724         map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
 725         if (!map) {
 726                 map = dso__new_map(event->ksymbol.name);
 727                 if (!map)
 728                         return -ENOMEM;
 729
 730                 map->start = event->ksymbol.addr;
 731                 map->end = map->start + event->ksymbol.len;
 732                 map_groups__insert(&machine->kmaps, map);
 733         }
 734
 735         sym = symbol__new(map->map_ip(map, map->start),
 736                           event->ksymbol.len,
 737                           0, 0, event->ksymbol.name);
 738         if (!sym)
 739                 return -ENOMEM;
 740         dso__insert_symbol(map->dso, sym);
 741         return 0;
 742 }
 743
 744 static int machine__process_ksymbol_unregister(struct machine *machine,
 745                                                union perf_event *event,
 746                                                struct perf_sample *sample __maybe_unused)
 747 {
 748         struct map *map;
 749
 750         map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
 751         if (map)
 752                 map_groups__remove(&machine->kmaps, map);
 753
 754         return 0;
 755 }
 756
 757 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 758                              union perf_event *event,
 759                              struct perf_sample *sample)
 760 {
 761         if (dump_trace)
 762                 perf_event__fprintf_ksymbol(event, stdout);
 763
 764         if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 765                 return machine__process_ksymbol_unregister(machine, event,
 766                                                            sample);
 767         return machine__process_ksymbol_register(machine, event, sample);
 768 }
 769
 770 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
 771 {
 772         const char *dup_filename;
 773
 774         if (!filename || !dso || !dso->long_name)
 775                 return;
 776         if (dso->long_name[0] != '[')
 777                 return;
 778         if (!strchr(filename, '/'))
 779                 return;
 780
 781         dup_filename = strdup(filename);
 782         if (!dup_filename)
 783                 return;
 784
 785         dso__set_long_name(dso, dup_filename, true);
 786 }
 787
 788 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 789                                         const char *filename)
 790 {
 791         struct map *map = NULL;
 792         struct dso *dso = NULL;
 793         struct kmod_path m;
 794
 795         if (kmod_path__parse_name(&m, filename))
 796                 return NULL;
 797
 798         map = map_groups__find_by_name(&machine->kmaps, m.name);
 799         if (map) {
 800                 /*
 801                  * If the map's dso is an offline module, give dso__load()
 802                  * a chance to find the file path of that module by fixing
 803                  * long_name.
 804                  */
 805                 dso__adjust_kmod_long_name(map->dso, filename);
 806                 goto out;
 807         }
 808
 809         dso = machine__findnew_module_dso(machine, &m, filename);
 810         if (dso == NULL)
 811                 goto out;
 812
 813         map = map__new2(start, dso);
 814         if (map == NULL)
 815                 goto out;
 816
 817         map_groups__insert(&machine->kmaps, map);
 818
 819         /* Put the map here because map_groups__insert alread got it */
 820         map__put(map);
 821 out:
 822         /* put the dso here, corresponding to  machine__findnew_module_dso */
 823         dso__put(dso);
 824         zfree(&m.name);
 825         return map;
 826 }
 827
 828 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 829 {
 830         struct rb_node *nd;
 831         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 832
 833         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 834                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 835                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 836         }
 837
 838         return ret;
 839 }
 840
 841 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 842                                      bool (skip)(struct dso *dso, int parm), int parm)
 843 {
 844         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 845 }
 846
 847 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 848                                      bool (skip)(struct dso *dso, int parm), int parm)
 849 {
 850         struct rb_node *nd;
 851         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 852
 853         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 854                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 855                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 856         }
 857         return ret;
 858 }
 859
 860 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 861 {
 862         int i;
 863         size_t printed = 0;
 864         struct dso *kdso = machine__kernel_map(machine)->dso;
 865
 866         if (kdso->has_build_id) {
 867                 char filename[PATH_MAX];
 868                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
 869                                            false))
 870                         printed += fprintf(fp, "[0] %s\n", filename);
 871         }
 872
 873         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 874                 printed += fprintf(fp, "[%d] %s\n",
 875                                    i + kdso->has_build_id, vmlinux_path[i]);
 876
 877         return printed;
 878 }
 879
 880 size_t machine__fprintf(struct machine *machine, FILE *fp)
 881 {
 882         struct rb_node *nd;
 883         size_t ret;
 884         int i;
 885
 886         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 887                 struct threads *threads = &machine->threads[i];
 888
 889                 down_read(&threads->lock);
 890
 891                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
 892
 893                 for (nd = rb_first_cached(&threads->entries); nd;
 894                      nd = rb_next(nd)) {
 895                         struct thread *pos = rb_entry(nd, struct thread, rb_node);
 896
 897                         ret += thread__fprintf(pos, fp);
 898                 }
 899
 900                 up_read(&threads->lock);
 901         }
 902         return ret;
 903 }
 904
 905 static struct dso *machine__get_kernel(struct machine *machine)
 906 {
 907         const char *vmlinux_name = machine->mmap_name;
 908         struct dso *kernel;
 909
 910         if (machine__is_host(machine)) {
 911                 if (symbol_conf.vmlinux_name)
 912                         vmlinux_name = symbol_conf.vmlinux_name;
 913
 914                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 915                                                  "[kernel]", DSO_TYPE_KERNEL);
 916         } else {
 917                 if (symbol_conf.default_guest_vmlinux_name)
 918                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 919
 920                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 921                                                  "[guest.kernel]",
 922                                                  DSO_TYPE_GUEST_KERNEL);
 923         }
 924
 925         if (kernel != NULL && (!kernel->has_build_id))
 926                 dso__read_running_kernel_build_id(kernel, machine);
 927
 928         return kernel;
 929 }
 930
 931 struct process_args {
 932         u64 start;
 933 };
 934
 935 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 936                                     size_t bufsz)
 937 {
 938         if (machine__is_default_guest(machine))
 939                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 940         else
 941                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 942 }
 943
 944 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 945
 946 /* Figure out the start address of kernel map from /proc/kallsyms.
 947  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 948  * symbol_name if it's not that important.
 949  */
 950 static int machine__get_running_kernel_start(struct machine *machine,
 951                                              const char **symbol_name,
 952                                              u64 *start, u64 *end)
 953 {
 954         char filename[PATH_MAX];
 955         int i, err = -1;
 956         const char *name;
 957         u64 addr = 0;
 958
 959         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 960
 961         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 962                 return 0;
 963
 964         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 965                 err = kallsyms__get_function_start(filename, name, &addr);
 966                 if (!err)
 967                         break;
 968         }
 969
 970         if (err)
 971                 return -1;
 972
 973         if (symbol_name)
 974                 *symbol_name = name;
 975
 976         *start = addr;
 977
 978         err = kallsyms__get_function_start(filename, "_etext", &addr);
 979         if (!err)
 980                 *end = addr;
 981
 982         return 0;
 983 }
 984
 985 int machine__create_extra_kernel_map(struct machine *machine,
 986                                      struct dso *kernel,
 987                                      struct extra_kernel_map *xm)
 988 {
 989         struct kmap *kmap;
 990         struct map *map;
 991
 992         map = map__new2(xm->start, kernel);
 993         if (!map)
 994                 return -1;
 995
 996         map->end   = xm->end;
 997         map->pgoff = xm->pgoff;
 998
 999         kmap = map__kmap(map);
1000
1001         kmap->kmaps = &machine->kmaps;
1002         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1003
1004         map_groups__insert(&machine->kmaps, map);
1005
1006         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1007                   kmap->name, map->start, map->end);
1008
1009         map__put(map);
1010
1011         return 0;
1012 }
1013
1014 static u64 find_entry_trampoline(struct dso *dso)
1015 {
1016         /* Duplicates are removed so lookup all aliases */
1017         const char *syms[] = {
1018                 "_entry_trampoline",
1019                 "__entry_trampoline_start",
1020                 "entry_SYSCALL_64_trampoline",
1021         };
1022         struct symbol *sym = dso__first_symbol(dso);
1023         unsigned int i;
1024
1025         for (; sym; sym = dso__next_symbol(sym)) {
1026                 if (sym->binding != STB_GLOBAL)
1027                         continue;
1028                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1029                         if (!strcmp(sym->name, syms[i]))
1030                                 return sym->start;
1031                 }
1032         }
1033
1034         return 0;
1035 }
1036
1037 /*
1038  * These values can be used for kernels that do not have symbols for the entry
1039  * trampolines in kallsyms.
1040  */
1041 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1042 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1043 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1044
1045 /* Map x86_64 PTI entry trampolines */
1046 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1047                                           struct dso *kernel)
1048 {
1049         struct map_groups *kmaps = &machine->kmaps;
1050         struct maps *maps = &kmaps->maps;
1051         int nr_cpus_avail, cpu;
1052         bool found = false;
1053         struct map *map;
1054         u64 pgoff;
1055
1056         /*
1057          * In the vmlinux case, pgoff is a virtual address which must now be
1058          * mapped to a vmlinux offset.
1059          */
1060         for (map = maps__first(maps); map; map = map__next(map)) {
1061                 struct kmap *kmap = __map__kmap(map);
1062                 struct map *dest_map;
1063
1064                 if (!kmap || !is_entry_trampoline(kmap->name))
1065                         continue;
1066
1067                 dest_map = map_groups__find(kmaps, map->pgoff);
1068                 if (dest_map != map)
1069                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1070                 found = true;
1071         }
1072         if (found || machine->trampolines_mapped)
1073                 return 0;
1074
1075         pgoff = find_entry_trampoline(kernel);
1076         if (!pgoff)
1077                 return 0;
1078
1079         nr_cpus_avail = machine__nr_cpus_avail(machine);
1080
1081         /* Add a 1 page map for each CPU's entry trampoline */
1082         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1083                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1084                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1085                          X86_64_ENTRY_TRAMPOLINE;
1086                 struct extra_kernel_map xm = {
1087                         .start = va,
1088                         .end   = va + page_size,
1089                         .pgoff = pgoff,
1090                 };
1091
1092                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1093
1094                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1095                         return -1;
1096         }
1097
1098         machine->trampolines_mapped = nr_cpus_avail;
1099
1100         return 0;
1101 }
1102
1103 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1104                                              struct dso *kernel __maybe_unused)
1105 {
1106         return 0;
1107 }
1108
1109 static int
1110 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1111 {
1112         struct kmap *kmap;
1113         struct map *map;
1114
1115         /* In case of renewal the kernel map, destroy previous one */
1116         machine__destroy_kernel_maps(machine);
1117
1118         machine->vmlinux_map = map__new2(0, kernel);
1119         if (machine->vmlinux_map == NULL)
1120                 return -1;
1121
1122         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1123         map = machine__kernel_map(machine);
1124         kmap = map__kmap(map);
1125         if (!kmap)
1126                 return -1;
1127
1128         kmap->kmaps = &machine->kmaps;
1129         map_groups__insert(&machine->kmaps, map);
1130
1131         return 0;
1132 }
1133
1134 void machine__destroy_kernel_maps(struct machine *machine)
1135 {
1136         struct kmap *kmap;
1137         struct map *map = machine__kernel_map(machine);
1138
1139         if (map == NULL)
1140                 return;
1141
1142         kmap = map__kmap(map);
1143         map_groups__remove(&machine->kmaps, map);
1144         if (kmap && kmap->ref_reloc_sym) {
1145                 zfree((char **)&kmap->ref_reloc_sym->name);
1146                 zfree(&kmap->ref_reloc_sym);
1147         }
1148
1149         map__zput(machine->vmlinux_map);
1150 }
1151
1152 int machines__create_guest_kernel_maps(struct machines *machines)
1153 {
1154         int ret = 0;
1155         struct dirent **namelist = NULL;
1156         int i, items = 0;
1157         char path[PATH_MAX];
1158         pid_t pid;
1159         char *endp;
1160
1161         if (symbol_conf.default_guest_vmlinux_name ||
1162             symbol_conf.default_guest_modules ||
1163             symbol_conf.default_guest_kallsyms) {
1164                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1165         }
1166
1167         if (symbol_conf.guestmount) {
1168                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1169                 if (items <= 0)
1170                         return -ENOENT;
1171                 for (i = 0; i < items; i++) {
1172                         if (!isdigit(namelist[i]->d_name[0])) {
1173                                 /* Filter out . and .. */
1174                                 continue;
1175                         }
1176                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1177                         if ((*endp != '\0') ||
1178                             (endp == namelist[i]->d_name) ||
1179                             (errno == ERANGE)) {
1180                                 pr_debug("invalid directory (%s). Skipping.\n",
1181                                          namelist[i]->d_name);
1182                                 continue;
1183                         }
1184                         sprintf(path, "%s/%s/proc/kallsyms",
1185                                 symbol_conf.guestmount,
1186                                 namelist[i]->d_name);
1187                         ret = access(path, R_OK);
1188                         if (ret) {
1189                                 pr_debug("Can't access file %s\n", path);
1190                                 goto failure;
1191                         }
1192                         machines__create_kernel_maps(machines, pid);
1193                 }
1194 failure:
1195                 free(namelist);
1196         }
1197
1198         return ret;
1199 }
1200
1201 void machines__destroy_kernel_maps(struct machines *machines)
1202 {
1203         struct rb_node *next = rb_first_cached(&machines->guests);
1204
1205         machine__destroy_kernel_maps(&machines->host);
1206
1207         while (next) {
1208                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1209
1210                 next = rb_next(&pos->rb_node);
1211                 rb_erase_cached(&pos->rb_node, &machines->guests);
1212                 machine__delete(pos);
1213         }
1214 }
1215
1216 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1217 {
1218         struct machine *machine = machines__findnew(machines, pid);
1219
1220         if (machine == NULL)
1221                 return -1;
1222
1223         return machine__create_kernel_maps(machine);
1224 }
1225
1226 int machine__load_kallsyms(struct machine *machine, const char *filename)
1227 {
1228         struct map *map = machine__kernel_map(machine);
1229         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1230
1231         if (ret > 0) {
1232                 dso__set_loaded(map->dso);
1233                 /*
1234                  * Since /proc/kallsyms will have multiple sessions for the
1235                  * kernel, with modules between them, fixup the end of all
1236                  * sections.
1237                  */
1238                 map_groups__fixup_end(&machine->kmaps);
1239         }
1240
1241         return ret;
1242 }
1243
1244 int machine__load_vmlinux_path(struct machine *machine)
1245 {
1246         struct map *map = machine__kernel_map(machine);
1247         int ret = dso__load_vmlinux_path(map->dso, map);
1248
1249         if (ret > 0)
1250                 dso__set_loaded(map->dso);
1251
1252         return ret;
1253 }
1254
1255 static char *get_kernel_version(const char *root_dir)
1256 {
1257         char version[PATH_MAX];
1258         FILE *file;
1259         char *name, *tmp;
1260         const char *prefix = "Linux version ";
1261
1262         sprintf(version, "%s/proc/version", root_dir);
1263         file = fopen(version, "r");
1264         if (!file)
1265                 return NULL;
1266
1267         tmp = fgets(version, sizeof(version), file);
1268         fclose(file);
1269         if (!tmp)
1270                 return NULL;
1271
1272         name = strstr(version, prefix);
1273         if (!name)
1274                 return NULL;
1275         name += strlen(prefix);
1276         tmp = strchr(name, ' ');
1277         if (tmp)
1278                 *tmp = '\0';
1279
1280         return strdup(name);
1281 }
1282
1283 static bool is_kmod_dso(struct dso *dso)
1284 {
1285         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1286                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1287 }
1288
1289 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1290                                        struct kmod_path *m)
1291 {
1292         char *long_name;
1293         struct map *map = map_groups__find_by_name(mg, m->name);
1294
1295         if (map == NULL)
1296                 return 0;
1297
1298         long_name = strdup(path);
1299         if (long_name == NULL)
1300                 return -ENOMEM;
1301
1302         dso__set_long_name(map->dso, long_name, true);
1303         dso__kernel_module_get_build_id(map->dso, "");
1304
1305         /*
1306          * Full name could reveal us kmod compression, so
1307          * we need to update the symtab_type if needed.
1308          */
1309         if (m->comp && is_kmod_dso(map->dso)) {
1310                 map->dso->symtab_type++;
1311                 map->dso->comp = m->comp;
1312         }
1313
1314         return 0;
1315 }
1316
1317 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1318                                 const char *dir_name, int depth)
1319 {
1320         struct dirent *dent;
1321         DIR *dir = opendir(dir_name);
1322         int ret = 0;
1323
1324         if (!dir) {
1325                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1326                 return -1;
1327         }
1328
1329         while ((dent = readdir(dir)) != NULL) {
1330                 char path[PATH_MAX];
1331                 struct stat st;
1332
1333                 /*sshfs might return bad dent->d_type, so we have to stat*/
1334                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1335                 if (stat(path, &st))
1336                         continue;
1337
1338                 if (S_ISDIR(st.st_mode)) {
1339                         if (!strcmp(dent->d_name, ".") ||
1340                             !strcmp(dent->d_name, ".."))
1341                                 continue;
1342
1343                         /* Do not follow top-level source and build symlinks */
1344                         if (depth == 0) {
1345                                 if (!strcmp(dent->d_name, "source") ||
1346                                     !strcmp(dent->d_name, "build"))
1347                                         continue;
1348                         }
1349
1350                         ret = map_groups__set_modules_path_dir(mg, path,
1351                                                                depth + 1);
1352                         if (ret < 0)
1353                                 goto out;
1354                 } else {
1355                         struct kmod_path m;
1356
1357                         ret = kmod_path__parse_name(&m, dent->d_name);
1358                         if (ret)
1359                                 goto out;
1360
1361                         if (m.kmod)
1362                                 ret = map_groups__set_module_path(mg, path, &m);
1363
1364                         zfree(&m.name);
1365
1366                         if (ret)
1367                                 goto out;
1368                 }
1369         }
1370
1371 out:
1372         closedir(dir);
1373         return ret;
1374 }
1375
1376 static int machine__set_modules_path(struct machine *machine)
1377 {
1378         char *version;
1379         char modules_path[PATH_MAX];
1380
1381         version = get_kernel_version(machine->root_dir);
1382         if (!version)
1383                 return -1;
1384
1385         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1386                  machine->root_dir, version);
1387         free(version);
1388
1389         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1390 }
1391 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1392                                 u64 *size __maybe_unused,
1393                                 const char *name __maybe_unused)
1394 {
1395         return 0;
1396 }
1397
1398 static int machine__create_module(void *arg, const char *name, u64 start,
1399                                   u64 size)
1400 {
1401         struct machine *machine = arg;
1402         struct map *map;
1403
1404         if (arch__fix_module_text_start(&start, &size, name) < 0)
1405                 return -1;
1406
1407         map = machine__findnew_module_map(machine, start, name);
1408         if (map == NULL)
1409                 return -1;
1410         map->end = start + size;
1411
1412         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1413
1414         return 0;
1415 }
1416
1417 static int machine__create_modules(struct machine *machine)
1418 {
1419         const char *modules;
1420         char path[PATH_MAX];
1421
1422         if (machine__is_default_guest(machine)) {
1423                 modules = symbol_conf.default_guest_modules;
1424         } else {
1425                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1426                 modules = path;
1427         }
1428
1429         if (symbol__restricted_filename(modules, "/proc/modules"))
1430                 return -1;
1431
1432         if (modules__parse(modules, machine, machine__create_module))
1433                 return -1;
1434
1435         if (!machine__set_modules_path(machine))
1436                 return 0;
1437
1438         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1439
1440         return 0;
1441 }
1442
1443 static void machine__set_kernel_mmap(struct machine *machine,
1444                                      u64 start, u64 end)
1445 {
1446         machine->vmlinux_map->start = start;
1447         machine->vmlinux_map->end   = end;
1448         /*
1449          * Be a bit paranoid here, some perf.data file came with
1450          * a zero sized synthesized MMAP event for the kernel.
1451          */
1452         if (start == 0 && end == 0)
1453                 machine->vmlinux_map->end = ~0ULL;
1454 }
1455
1456 static void machine__update_kernel_mmap(struct machine *machine,
1457                                      u64 start, u64 end)
1458 {
1459         struct map *map = machine__kernel_map(machine);
1460
1461         map__get(map);
1462         map_groups__remove(&machine->kmaps, map);
1463
1464         machine__set_kernel_mmap(machine, start, end);
1465
1466         map_groups__insert(&machine->kmaps, map);
1467         map__put(map);
1468 }
1469
1470 int machine__create_kernel_maps(struct machine *machine)
1471 {
1472         struct dso *kernel = machine__get_kernel(machine);
1473         const char *name = NULL;
1474         struct map *map;
1475         u64 start = 0, end = ~0ULL;
1476         int ret;
1477
1478         if (kernel == NULL)
1479                 return -1;
1480
1481         ret = __machine__create_kernel_maps(machine, kernel);
1482         if (ret < 0)
1483                 goto out_put;
1484
1485         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1486                 if (machine__is_host(machine))
1487                         pr_debug("Problems creating module maps, "
1488                                  "continuing anyway...\n");
1489                 else
1490                         pr_debug("Problems creating module maps for guest %d, "
1491                                  "continuing anyway...\n", machine->pid);
1492         }
1493
1494         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1495                 if (name &&
1496                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1497                         machine__destroy_kernel_maps(machine);
1498                         ret = -1;
1499                         goto out_put;
1500                 }
1501
1502                 /*
1503                  * we have a real start address now, so re-order the kmaps
1504                  * assume it's the last in the kmaps
1505                  */
1506                 machine__update_kernel_mmap(machine, start, end);
1507         }
1508
1509         if (machine__create_extra_kernel_maps(machine, kernel))
1510                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1511
1512         if (end == ~0ULL) {
1513                 /* update end address of the kernel map using adjacent module address */
1514                 map = map__next(machine__kernel_map(machine));
1515                 if (map)
1516                         machine__set_kernel_mmap(machine, start, map->start);
1517         }
1518
1519 out_put:
1520         dso__put(kernel);
1521         return ret;
1522 }
1523
1524 static bool machine__uses_kcore(struct machine *machine)
1525 {
1526         struct dso *dso;
1527
1528         list_for_each_entry(dso, &machine->dsos.head, node) {
1529                 if (dso__is_kcore(dso))
1530                         return true;
1531         }
1532
1533         return false;
1534 }
1535
1536 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1537                                              union perf_event *event)
1538 {
1539         return machine__is(machine, "x86_64") &&
1540                is_entry_trampoline(event->mmap.filename);
1541 }
1542
1543 static int machine__process_extra_kernel_map(struct machine *machine,
1544                                              union perf_event *event)
1545 {
1546         struct map *kernel_map = machine__kernel_map(machine);
1547         struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1548         struct extra_kernel_map xm = {
1549                 .start = event->mmap.start,
1550                 .end   = event->mmap.start + event->mmap.len,
1551                 .pgoff = event->mmap.pgoff,
1552         };
1553
1554         if (kernel == NULL)
1555                 return -1;
1556
1557         strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1558
1559         return machine__create_extra_kernel_map(machine, kernel, &xm);
1560 }
1561
1562 static int machine__process_kernel_mmap_event(struct machine *machine,
1563                                               union perf_event *event)
1564 {
1565         struct map *map;
1566         enum dso_kernel_type kernel_type;
1567         bool is_kernel_mmap;
1568
1569         /* If we have maps from kcore then we do not need or want any others */
1570         if (machine__uses_kcore(machine))
1571                 return 0;
1572
1573         if (machine__is_host(machine))
1574                 kernel_type = DSO_TYPE_KERNEL;
1575         else
1576                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1577
1578         is_kernel_mmap = memcmp(event->mmap.filename,
1579                                 machine->mmap_name,
1580                                 strlen(machine->mmap_name) - 1) == 0;
1581         if (event->mmap.filename[0] == '/' ||
1582             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1583                 map = machine__findnew_module_map(machine, event->mmap.start,
1584                                                   event->mmap.filename);
1585                 if (map == NULL)
1586                         goto out_problem;
1587
1588                 map->end = map->start + event->mmap.len;
1589         } else if (is_kernel_mmap) {
1590                 const char *symbol_name = (event->mmap.filename +
1591                                 strlen(machine->mmap_name));
1592                 /*
1593                  * Should be there already, from the build-id table in
1594                  * the header.
1595                  */
1596                 struct dso *kernel = NULL;
1597                 struct dso *dso;
1598
1599                 down_read(&machine->dsos.lock);
1600
1601                 list_for_each_entry(dso, &machine->dsos.head, node) {
1602
1603                         /*
1604                          * The cpumode passed to is_kernel_module is not the
1605                          * cpumode of *this* event. If we insist on passing
1606                          * correct cpumode to is_kernel_module, we should
1607                          * record the cpumode when we adding this dso to the
1608                          * linked list.
1609                          *
1610                          * However we don't really need passing correct
1611                          * cpumode.  We know the correct cpumode must be kernel
1612                          * mode (if not, we should not link it onto kernel_dsos
1613                          * list).
1614                          *
1615                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1616                          * is_kernel_module() treats it as a kernel cpumode.
1617                          */
1618
1619                         if (!dso->kernel ||
1620                             is_kernel_module(dso->long_name,
1621                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1622                                 continue;
1623
1624
1625                         kernel = dso;
1626                         break;
1627                 }
1628
1629                 up_read(&machine->dsos.lock);
1630
1631                 if (kernel == NULL)
1632                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1633                 if (kernel == NULL)
1634                         goto out_problem;
1635
1636                 kernel->kernel = kernel_type;
1637                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1638                         dso__put(kernel);
1639                         goto out_problem;
1640                 }
1641
1642                 if (strstr(kernel->long_name, "vmlinux"))
1643                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1644
1645                 machine__update_kernel_mmap(machine, event->mmap.start,
1646                                          event->mmap.start + event->mmap.len);
1647
1648                 /*
1649                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1650                  * symbol. Effectively having zero here means that at record
1651                  * time /proc/sys/kernel/kptr_restrict was non zero.
1652                  */
1653                 if (event->mmap.pgoff != 0) {
1654                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1655                                                         symbol_name,
1656                                                         event->mmap.pgoff);
1657                 }
1658
1659                 if (machine__is_default_guest(machine)) {
1660                         /*
1661                          * preload dso of guest kernel and modules
1662                          */
1663                         dso__load(kernel, machine__kernel_map(machine));
1664                 }
1665         } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1666                 return machine__process_extra_kernel_map(machine, event);
1667         }
1668         return 0;
1669 out_problem:
1670         return -1;
1671 }
1672
1673 int machine__process_mmap2_event(struct machine *machine,
1674                                  union perf_event *event,
1675                                  struct perf_sample *sample)
1676 {
1677         struct thread *thread;
1678         struct map *map;
1679         int ret = 0;
1680
1681         if (dump_trace)
1682                 perf_event__fprintf_mmap2(event, stdout);
1683
1684         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1685             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1686                 ret = machine__process_kernel_mmap_event(machine, event);
1687                 if (ret < 0)
1688                         goto out_problem;
1689                 return 0;
1690         }
1691
1692         thread = machine__findnew_thread(machine, event->mmap2.pid,
1693                                         event->mmap2.tid);
1694         if (thread == NULL)
1695                 goto out_problem;
1696
1697         map = map__new(machine, event->mmap2.start,
1698                         event->mmap2.len, event->mmap2.pgoff,
1699                         event->mmap2.maj,
1700                         event->mmap2.min, event->mmap2.ino,
1701                         event->mmap2.ino_generation,
1702                         event->mmap2.prot,
1703                         event->mmap2.flags,
1704                         event->mmap2.filename, thread);
1705
1706         if (map == NULL)
1707                 goto out_problem_map;
1708
1709         ret = thread__insert_map(thread, map);
1710         if (ret)
1711                 goto out_problem_insert;
1712
1713         thread__put(thread);
1714         map__put(map);
1715         return 0;
1716
1717 out_problem_insert:
1718         map__put(map);
1719 out_problem_map:
1720         thread__put(thread);
1721 out_problem:
1722         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1723         return 0;
1724 }
1725
1726 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1727                                 struct perf_sample *sample)
1728 {
1729         struct thread *thread;
1730         struct map *map;
1731         u32 prot = 0;
1732         int ret = 0;
1733
1734         if (dump_trace)
1735                 perf_event__fprintf_mmap(event, stdout);
1736
1737         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1738             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1739                 ret = machine__process_kernel_mmap_event(machine, event);
1740                 if (ret < 0)
1741                         goto out_problem;
1742                 return 0;
1743         }
1744
1745         thread = machine__findnew_thread(machine, event->mmap.pid,
1746                                          event->mmap.tid);
1747         if (thread == NULL)
1748                 goto out_problem;
1749
1750         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1751                 prot = PROT_EXEC;
1752
1753         map = map__new(machine, event->mmap.start,
1754                         event->mmap.len, event->mmap.pgoff,
1755                         0, 0, 0, 0, prot, 0,
1756                         event->mmap.filename,
1757                         thread);
1758
1759         if (map == NULL)
1760                 goto out_problem_map;
1761
1762         ret = thread__insert_map(thread, map);
1763         if (ret)
1764                 goto out_problem_insert;
1765
1766         thread__put(thread);
1767         map__put(map);
1768         return 0;
1769
1770 out_problem_insert:
1771         map__put(map);
1772 out_problem_map:
1773         thread__put(thread);
1774 out_problem:
1775         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1776         return 0;
1777 }
1778
1779 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1780 {
1781         struct threads *threads = machine__threads(machine, th->tid);
1782
1783         if (threads->last_match == th)
1784                 threads__set_last_match(threads, NULL);
1785
1786         if (lock)
1787                 down_write(&threads->lock);
1788
1789         BUG_ON(refcount_read(&th->refcnt) == 0);
1790
1791         rb_erase_cached(&th->rb_node, &threads->entries);
1792         RB_CLEAR_NODE(&th->rb_node);
1793         --threads->nr;
1794         /*
1795          * Move it first to the dead_threads list, then drop the reference,
1796          * if this is the last reference, then the thread__delete destructor
1797          * will be called and we will remove it from the dead_threads list.
1798          */
1799         list_add_tail(&th->node, &threads->dead);
1800
1801         /*
1802          * We need to do the put here because if this is the last refcount,
1803          * then we will be touching the threads->dead head when removing the
1804          * thread.
1805          */
1806         thread__put(th);
1807
1808         if (lock)
1809                 up_write(&threads->lock);
1810 }
1811
1812 void machine__remove_thread(struct machine *machine, struct thread *th)
1813 {
1814         return __machine__remove_thread(machine, th, true);
1815 }
1816
1817 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1818                                 struct perf_sample *sample)
1819 {
1820         struct thread *thread = machine__find_thread(machine,
1821                                                      event->fork.pid,
1822                                                      event->fork.tid);
1823         struct thread *parent = machine__findnew_thread(machine,
1824                                                         event->fork.ppid,
1825                                                         event->fork.ptid);
1826         bool do_maps_clone = true;
1827         int err = 0;
1828
1829         if (dump_trace)
1830                 perf_event__fprintf_task(event, stdout);
1831
1832         /*
1833          * There may be an existing thread that is not actually the parent,
1834          * either because we are processing events out of order, or because the
1835          * (fork) event that would have removed the thread was lost. Assume the
1836          * latter case and continue on as best we can.
1837          */
1838         if (parent->pid_ != (pid_t)event->fork.ppid) {
1839                 dump_printf("removing erroneous parent thread %d/%d\n",
1840                             parent->pid_, parent->tid);
1841                 machine__remove_thread(machine, parent);
1842                 thread__put(parent);
1843                 parent = machine__findnew_thread(machine, event->fork.ppid,
1844                                                  event->fork.ptid);
1845         }
1846
1847         /* if a thread currently exists for the thread id remove it */
1848         if (thread != NULL) {
1849                 machine__remove_thread(machine, thread);
1850                 thread__put(thread);
1851         }
1852
1853         thread = machine__findnew_thread(machine, event->fork.pid,
1854                                          event->fork.tid);
1855         /*
1856          * When synthesizing FORK events, we are trying to create thread
1857          * objects for the already running tasks on the machine.
1858          *
1859          * Normally, for a kernel FORK event, we want to clone the parent's
1860          * maps because that is what the kernel just did.
1861          *
1862          * But when synthesizing, this should not be done.  If we do, we end up
1863          * with overlapping maps as we process the sythesized MMAP2 events that
1864          * get delivered shortly thereafter.
1865          *
1866          * Use the FORK event misc flags in an internal way to signal this
1867          * situation, so we can elide the map clone when appropriate.
1868          */
1869         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1870                 do_maps_clone = false;
1871
1872         if (thread == NULL || parent == NULL ||
1873             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1874                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1875                 err = -1;
1876         }
1877         thread__put(thread);
1878         thread__put(parent);
1879
1880         return err;
1881 }
1882
1883 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1884                                 struct perf_sample *sample __maybe_unused)
1885 {
1886         struct thread *thread = machine__find_thread(machine,
1887                                                      event->fork.pid,
1888                                                      event->fork.tid);
1889
1890         if (dump_trace)
1891                 perf_event__fprintf_task(event, stdout);
1892
1893         if (thread != NULL) {
1894                 thread__exited(thread);
1895                 thread__put(thread);
1896         }
1897
1898         return 0;
1899 }
1900
1901 int machine__process_event(struct machine *machine, union perf_event *event,
1902                            struct perf_sample *sample)
1903 {
1904         int ret;
1905
1906         switch (event->header.type) {
1907         case PERF_RECORD_COMM:
1908                 ret = machine__process_comm_event(machine, event, sample); break;
1909         case PERF_RECORD_MMAP:
1910                 ret = machine__process_mmap_event(machine, event, sample); break;
1911         case PERF_RECORD_NAMESPACES:
1912                 ret = machine__process_namespaces_event(machine, event, sample); break;
1913         case PERF_RECORD_MMAP2:
1914                 ret = machine__process_mmap2_event(machine, event, sample); break;
1915         case PERF_RECORD_FORK:
1916                 ret = machine__process_fork_event(machine, event, sample); break;
1917         case PERF_RECORD_EXIT:
1918                 ret = machine__process_exit_event(machine, event, sample); break;
1919         case PERF_RECORD_LOST:
1920                 ret = machine__process_lost_event(machine, event, sample); break;
1921         case PERF_RECORD_AUX:
1922                 ret = machine__process_aux_event(machine, event); break;
1923         case PERF_RECORD_ITRACE_START:
1924                 ret = machine__process_itrace_start_event(machine, event); break;
1925         case PERF_RECORD_LOST_SAMPLES:
1926                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1927         case PERF_RECORD_SWITCH:
1928         case PERF_RECORD_SWITCH_CPU_WIDE:
1929                 ret = machine__process_switch_event(machine, event); break;
1930         case PERF_RECORD_KSYMBOL:
1931                 ret = machine__process_ksymbol(machine, event, sample); break;
1932         case PERF_RECORD_BPF_EVENT:
1933                 ret = machine__process_bpf(machine, event, sample); break;
1934         default:
1935                 ret = -1;
1936                 break;
1937         }
1938
1939         return ret;
1940 }
1941
1942 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1943 {
1944         if (!regexec(regex, sym->name, 0, NULL, 0))
1945                 return 1;
1946         return 0;
1947 }
1948
1949 static void ip__resolve_ams(struct thread *thread,
1950                             struct addr_map_symbol *ams,
1951                             u64 ip)
1952 {
1953         struct addr_location al;
1954
1955         memset(&al, 0, sizeof(al));
1956         /*
1957          * We cannot use the header.misc hint to determine whether a
1958          * branch stack address is user, kernel, guest, hypervisor.
1959          * Branches may straddle the kernel/user/hypervisor boundaries.
1960          * Thus, we have to try consecutively until we find a match
1961          * or else, the symbol is unknown
1962          */
1963         thread__find_cpumode_addr_location(thread, ip, &al);
1964
1965         ams->addr = ip;
1966         ams->al_addr = al.addr;
1967         ams->sym = al.sym;
1968         ams->map = al.map;
1969         ams->phys_addr = 0;
1970 }
1971
1972 static void ip__resolve_data(struct thread *thread,
1973                              u8 m, struct addr_map_symbol *ams,
1974                              u64 addr, u64 phys_addr)
1975 {
1976         struct addr_location al;
1977
1978         memset(&al, 0, sizeof(al));
1979
1980         thread__find_symbol(thread, m, addr, &al);
1981
1982         ams->addr = addr;
1983         ams->al_addr = al.addr;
1984         ams->sym = al.sym;
1985         ams->map = al.map;
1986         ams->phys_addr = phys_addr;
1987 }
1988
1989 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1990                                      struct addr_location *al)
1991 {
1992         struct mem_info *mi = mem_info__new();
1993
1994         if (!mi)
1995                 return NULL;
1996
1997         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1998         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1999                          sample->addr, sample->phys_addr);
2000         mi->data_src.val = sample->data_src;
2001
2002         return mi;
2003 }
2004
2005 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
2006 {
2007         char *srcline = NULL;
2008
2009         if (!map || callchain_param.key == CCKEY_FUNCTION)
2010                 return srcline;
2011
2012         srcline = srcline__tree_find(&map->dso->srclines, ip);
2013         if (!srcline) {
2014                 bool show_sym = false;
2015                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2016
2017                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2018                                       sym, show_sym, show_addr, ip);
2019                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2020         }
2021
2022         return srcline;
2023 }
2024
2025 struct iterations {
2026         int nr_loop_iter;
2027         u64 cycles;
2028 };
2029
2030 static int add_callchain_ip(struct thread *thread,
2031                             struct callchain_cursor *cursor,
2032                             struct symbol **parent,
2033                             struct addr_location *root_al,
2034                             u8 *cpumode,
2035                             u64 ip,
2036                             bool branch,
2037                             struct branch_flags *flags,
2038                             struct iterations *iter,
2039                             u64 branch_from)
2040 {
2041         struct addr_location al;
2042         int nr_loop_iter = 0;
2043         u64 iter_cycles = 0;
2044         const char *srcline = NULL;
2045
2046         al.filtered = 0;
2047         al.sym = NULL;
2048         if (!cpumode) {
2049                 thread__find_cpumode_addr_location(thread, ip, &al);
2050         } else {
2051                 if (ip >= PERF_CONTEXT_MAX) {
2052                         switch (ip) {
2053                         case PERF_CONTEXT_HV:
2054                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2055                                 break;
2056                         case PERF_CONTEXT_KERNEL:
2057                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2058                                 break;
2059                         case PERF_CONTEXT_USER:
2060                                 *cpumode = PERF_RECORD_MISC_USER;
2061                                 break;
2062                         default:
2063                                 pr_debug("invalid callchain context: "
2064                                          "%"PRId64"\n", (s64) ip);
2065                                 /*
2066                                  * It seems the callchain is corrupted.
2067                                  * Discard all.
2068                                  */
2069                                 callchain_cursor_reset(cursor);
2070                                 return 1;
2071                         }
2072                         return 0;
2073                 }
2074                 thread__find_symbol(thread, *cpumode, ip, &al);
2075         }
2076
2077         if (al.sym != NULL) {
2078                 if (perf_hpp_list.parent && !*parent &&
2079                     symbol__match_regex(al.sym, &parent_regex))
2080                         *parent = al.sym;
2081                 else if (have_ignore_callees && root_al &&
2082                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2083                         /* Treat this symbol as the root,
2084                            forgetting its callees. */
2085                         *root_al = al;
2086                         callchain_cursor_reset(cursor);
2087                 }
2088         }
2089
2090         if (symbol_conf.hide_unresolved && al.sym == NULL)
2091                 return 0;
2092
2093         if (iter) {
2094                 nr_loop_iter = iter->nr_loop_iter;
2095                 iter_cycles = iter->cycles;
2096         }
2097
2098         srcline = callchain_srcline(al.map, al.sym, al.addr);
2099         return callchain_cursor_append(cursor, ip, al.map, al.sym,
2100                                        branch, flags, nr_loop_iter,
2101                                        iter_cycles, branch_from, srcline);
2102 }
2103
2104 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2105                                            struct addr_location *al)
2106 {
2107         unsigned int i;
2108         const struct branch_stack *bs = sample->branch_stack;
2109         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2110
2111         if (!bi)
2112                 return NULL;
2113
2114         for (i = 0; i < bs->nr; i++) {
2115                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2116                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2117                 bi[i].flags = bs->entries[i].flags;
2118         }
2119         return bi;
2120 }
2121
2122 static void save_iterations(struct iterations *iter,
2123                             struct branch_entry *be, int nr)
2124 {
2125         int i;
2126
2127         iter->nr_loop_iter++;
2128         iter->cycles = 0;
2129
2130         for (i = 0; i < nr; i++)
2131                 iter->cycles += be[i].flags.cycles;
2132 }
2133
2134 #define CHASHSZ 127
2135 #define CHASHBITS 7
2136 #define NO_ENTRY 0xff
2137
2138 #define PERF_MAX_BRANCH_DEPTH 127
2139
2140 /* Remove loops. */
2141 static int remove_loops(struct branch_entry *l, int nr,
2142                         struct iterations *iter)
2143 {
2144         int i, j, off;
2145         unsigned char chash[CHASHSZ];
2146
2147         memset(chash, NO_ENTRY, sizeof(chash));
2148
2149         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2150
2151         for (i = 0; i < nr; i++) {
2152                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2153
2154                 /* no collision handling for now */
2155                 if (chash[h] == NO_ENTRY) {
2156                         chash[h] = i;
2157                 } else if (l[chash[h]].from == l[i].from) {
2158                         bool is_loop = true;
2159                         /* check if it is a real loop */
2160                         off = 0;
2161                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2162                                 if (l[j].from != l[i + off].from) {
2163                                         is_loop = false;
2164                                         break;
2165                                 }
2166                         if (is_loop) {
2167                                 j = nr - (i + off);
2168                                 if (j > 0) {
2169                                         save_iterations(iter + i + off,
2170                                                 l + i, off);
2171
2172                                         memmove(iter + i, iter + i + off,
2173                                                 j * sizeof(*iter));
2174
2175                                         memmove(l + i, l + i + off,
2176                                                 j * sizeof(*l));
2177                                 }
2178
2179                                 nr -= off;
2180                         }
2181                 }
2182         }
2183         return nr;
2184 }
2185
2186 /*
2187  * Recolve LBR callstack chain sample
2188  * Return:
2189  * 1 on success get LBR callchain information
2190  * 0 no available LBR callchain information, should try fp
2191  * negative error code on other errors.
2192  */
2193 static int resolve_lbr_callchain_sample(struct thread *thread,
2194                                         struct callchain_cursor *cursor,
2195                                         struct perf_sample *sample,
2196                                         struct symbol **parent,
2197                                         struct addr_location *root_al,
2198                                         int max_stack)
2199 {
2200         struct ip_callchain *chain = sample->callchain;
2201         int chain_nr = min(max_stack, (int)chain->nr), i;
2202         u8 cpumode = PERF_RECORD_MISC_USER;
2203         u64 ip, branch_from = 0;
2204
2205         for (i = 0; i < chain_nr; i++) {
2206                 if (chain->ips[i] == PERF_CONTEXT_USER)
2207                         break;
2208         }
2209
2210         /* LBR only affects the user callchain */
2211         if (i != chain_nr) {
2212                 struct branch_stack *lbr_stack = sample->branch_stack;
2213                 int lbr_nr = lbr_stack->nr, j, k;
2214                 bool branch;
2215                 struct branch_flags *flags;
2216                 /*
2217                  * LBR callstack can only get user call chain.
2218                  * The mix_chain_nr is kernel call chain
2219                  * number plus LBR user call chain number.
2220                  * i is kernel call chain number,
2221                  * 1 is PERF_CONTEXT_USER,
2222                  * lbr_nr + 1 is the user call chain number.
2223                  * For details, please refer to the comments
2224                  * in callchain__printf
2225                  */
2226                 int mix_chain_nr = i + 1 + lbr_nr + 1;
2227
2228                 for (j = 0; j < mix_chain_nr; j++) {
2229                         int err;
2230                         branch = false;
2231                         flags = NULL;
2232
2233                         if (callchain_param.order == ORDER_CALLEE) {
2234                                 if (j < i + 1)
2235                                         ip = chain->ips[j];
2236                                 else if (j > i + 1) {
2237                                         k = j - i - 2;
2238                                         ip = lbr_stack->entries[k].from;
2239                                         branch = true;
2240                                         flags = &lbr_stack->entries[k].flags;
2241                                 } else {
2242                                         ip = lbr_stack->entries[0].to;
2243                                         branch = true;
2244                                         flags = &lbr_stack->entries[0].flags;
2245                                         branch_from =
2246                                                 lbr_stack->entries[0].from;
2247                                 }
2248                         } else {
2249                                 if (j < lbr_nr) {
2250                                         k = lbr_nr - j - 1;
2251                                         ip = lbr_stack->entries[k].from;
2252                                         branch = true;
2253                                         flags = &lbr_stack->entries[k].flags;
2254                                 }
2255                                 else if (j > lbr_nr)
2256                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
2257                                 else {
2258                                         ip = lbr_stack->entries[0].to;
2259                                         branch = true;
2260                                         flags = &lbr_stack->entries[0].flags;
2261                                         branch_from =
2262                                                 lbr_stack->entries[0].from;
2263                                 }
2264                         }
2265
2266                         err = add_callchain_ip(thread, cursor, parent,
2267                                                root_al, &cpumode, ip,
2268                                                branch, flags, NULL,
2269                                                branch_from);
2270                         if (err)
2271                                 return (err < 0) ? err : 0;
2272                 }
2273                 return 1;
2274         }
2275
2276         return 0;
2277 }
2278
2279 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2280                              struct callchain_cursor *cursor,
2281                              struct symbol **parent,
2282                              struct addr_location *root_al,
2283                              u8 *cpumode, int ent)
2284 {
2285         int err = 0;
2286
2287         while (--ent >= 0) {
2288                 u64 ip = chain->ips[ent];
2289
2290                 if (ip >= PERF_CONTEXT_MAX) {
2291                         err = add_callchain_ip(thread, cursor, parent,
2292                                                root_al, cpumode, ip,
2293                                                false, NULL, NULL, 0);
2294                         break;
2295                 }
2296         }
2297         return err;
2298 }
2299
2300 static int thread__resolve_callchain_sample(struct thread *thread,
2301                                             struct callchain_cursor *cursor,
2302                                             struct evsel *evsel,
2303                                             struct perf_sample *sample,
2304                                             struct symbol **parent,
2305                                             struct addr_location *root_al,
2306                                             int max_stack)
2307 {
2308         struct branch_stack *branch = sample->branch_stack;
2309         struct ip_callchain *chain = sample->callchain;
2310         int chain_nr = 0;
2311         u8 cpumode = PERF_RECORD_MISC_USER;
2312         int i, j, err, nr_entries;
2313         int skip_idx = -1;
2314         int first_call = 0;
2315
2316         if (chain)
2317                 chain_nr = chain->nr;
2318
2319         if (perf_evsel__has_branch_callstack(evsel)) {
2320                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2321                                                    root_al, max_stack);
2322                 if (err)
2323                         return (err < 0) ? err : 0;
2324         }
2325
2326         /*
2327          * Based on DWARF debug information, some architectures skip
2328          * a callchain entry saved by the kernel.
2329          */
2330         skip_idx = arch_skip_callchain_idx(thread, chain);
2331
2332         /*
2333          * Add branches to call stack for easier browsing. This gives
2334          * more context for a sample than just the callers.
2335          *
2336          * This uses individual histograms of paths compared to the
2337          * aggregated histograms the normal LBR mode uses.
2338          *
2339          * Limitations for now:
2340          * - No extra filters
2341          * - No annotations (should annotate somehow)
2342          */
2343
2344         if (branch && callchain_param.branch_callstack) {
2345                 int nr = min(max_stack, (int)branch->nr);
2346                 struct branch_entry be[nr];
2347                 struct iterations iter[nr];
2348
2349                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2350                         pr_warning("corrupted branch chain. skipping...\n");
2351                         goto check_calls;
2352                 }
2353
2354                 for (i = 0; i < nr; i++) {
2355                         if (callchain_param.order == ORDER_CALLEE) {
2356                                 be[i] = branch->entries[i];
2357
2358                                 if (chain == NULL)
2359                                         continue;
2360
2361                                 /*
2362                                  * Check for overlap into the callchain.
2363                                  * The return address is one off compared to
2364                                  * the branch entry. To adjust for this
2365                                  * assume the calling instruction is not longer
2366                                  * than 8 bytes.
2367                                  */
2368                                 if (i == skip_idx ||
2369                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2370                                         first_call++;
2371                                 else if (be[i].from < chain->ips[first_call] &&
2372                                     be[i].from >= chain->ips[first_call] - 8)
2373                                         first_call++;
2374                         } else
2375                                 be[i] = branch->entries[branch->nr - i - 1];
2376                 }
2377
2378                 memset(iter, 0, sizeof(struct iterations) * nr);
2379                 nr = remove_loops(be, nr, iter);
2380
2381                 for (i = 0; i < nr; i++) {
2382                         err = add_callchain_ip(thread, cursor, parent,
2383                                                root_al,
2384                                                NULL, be[i].to,
2385                                                true, &be[i].flags,
2386                                                NULL, be[i].from);
2387
2388                         if (!err)
2389                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2390                                                        NULL, be[i].from,
2391                                                        true, &be[i].flags,
2392                                                        &iter[i], 0);
2393                         if (err == -EINVAL)
2394                                 break;
2395                         if (err)
2396                                 return err;
2397                 }
2398
2399                 if (chain_nr == 0)
2400                         return 0;
2401
2402                 chain_nr -= nr;
2403         }
2404
2405 check_calls:
2406         if (callchain_param.order != ORDER_CALLEE) {
2407                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2408                                         &cpumode, chain->nr - first_call);
2409                 if (err)
2410                         return (err < 0) ? err : 0;
2411         }
2412         for (i = first_call, nr_entries = 0;
2413              i < chain_nr && nr_entries < max_stack; i++) {
2414                 u64 ip;
2415
2416                 if (callchain_param.order == ORDER_CALLEE)
2417                         j = i;
2418                 else
2419                         j = chain->nr - i - 1;
2420
2421 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2422                 if (j == skip_idx)
2423                         continue;
2424 #endif
2425                 ip = chain->ips[j];
2426                 if (ip < PERF_CONTEXT_MAX)
2427                        ++nr_entries;
2428                 else if (callchain_param.order != ORDER_CALLEE) {
2429                         err = find_prev_cpumode(chain, thread, cursor, parent,
2430                                                 root_al, &cpumode, j);
2431                         if (err)
2432                                 return (err < 0) ? err : 0;
2433                         continue;
2434                 }
2435
2436                 err = add_callchain_ip(thread, cursor, parent,
2437                                        root_al, &cpumode, ip,
2438                                        false, NULL, NULL, 0);
2439
2440                 if (err)
2441                         return (err < 0) ? err : 0;
2442         }
2443
2444         return 0;
2445 }
2446
2447 static int append_inlines(struct callchain_cursor *cursor,
2448                           struct map *map, struct symbol *sym, u64 ip)
2449 {
2450         struct inline_node *inline_node;
2451         struct inline_list *ilist;
2452         u64 addr;
2453         int ret = 1;
2454
2455         if (!symbol_conf.inline_name || !map || !sym)
2456                 return ret;
2457
2458         addr = map__map_ip(map, ip);
2459         addr = map__rip_2objdump(map, addr);
2460
2461         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2462         if (!inline_node) {
2463                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2464                 if (!inline_node)
2465                         return ret;
2466                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2467         }
2468
2469         list_for_each_entry(ilist, &inline_node->val, list) {
2470                 ret = callchain_cursor_append(cursor, ip, map,
2471                                               ilist->symbol, false,
2472                                               NULL, 0, 0, 0, ilist->srcline);
2473
2474                 if (ret != 0)
2475                         return ret;
2476         }
2477
2478         return ret;
2479 }
2480
2481 static int unwind_entry(struct unwind_entry *entry, void *arg)
2482 {
2483         struct callchain_cursor *cursor = arg;
2484         const char *srcline = NULL;
2485         u64 addr = entry->ip;
2486
2487         if (symbol_conf.hide_unresolved && entry->sym == NULL)
2488                 return 0;
2489
2490         if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2491                 return 0;
2492
2493         /*
2494          * Convert entry->ip from a virtual address to an offset in
2495          * its corresponding binary.
2496          */
2497         if (entry->map)
2498                 addr = map__map_ip(entry->map, entry->ip);
2499
2500         srcline = callchain_srcline(entry->map, entry->sym, addr);
2501         return callchain_cursor_append(cursor, entry->ip,
2502                                        entry->map, entry->sym,
2503                                        false, NULL, 0, 0, 0, srcline);
2504 }
2505
2506 static int thread__resolve_callchain_unwind(struct thread *thread,
2507                                             struct callchain_cursor *cursor,
2508                                             struct evsel *evsel,
2509                                             struct perf_sample *sample,
2510                                             int max_stack)
2511 {
2512         /* Can we do dwarf post unwind? */
2513         if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2514               (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2515                 return 0;
2516
2517         /* Bail out if nothing was captured. */
2518         if ((!sample->user_regs.regs) ||
2519             (!sample->user_stack.size))
2520                 return 0;
2521
2522         return unwind__get_entries(unwind_entry, cursor,
2523                                    thread, sample, max_stack);
2524 }
2525
2526 int thread__resolve_callchain(struct thread *thread,
2527                               struct callchain_cursor *cursor,
2528                               struct evsel *evsel,
2529                               struct perf_sample *sample,
2530                               struct symbol **parent,
2531                               struct addr_location *root_al,
2532                               int max_stack)
2533 {
2534         int ret = 0;
2535
2536         callchain_cursor_reset(cursor);
2537
2538         if (callchain_param.order == ORDER_CALLEE) {
2539                 ret = thread__resolve_callchain_sample(thread, cursor,
2540                                                        evsel, sample,
2541                                                        parent, root_al,
2542                                                        max_stack);
2543                 if (ret)
2544                         return ret;
2545                 ret = thread__resolve_callchain_unwind(thread, cursor,
2546                                                        evsel, sample,
2547                                                        max_stack);
2548         } else {
2549                 ret = thread__resolve_callchain_unwind(thread, cursor,
2550                                                        evsel, sample,
2551                                                        max_stack);
2552                 if (ret)
2553                         return ret;
2554                 ret = thread__resolve_callchain_sample(thread, cursor,
2555                                                        evsel, sample,
2556                                                        parent, root_al,
2557                                                        max_stack);
2558         }
2559
2560         return ret;
2561 }
2562
2563 int machine__for_each_thread(struct machine *machine,
2564                              int (*fn)(struct thread *thread, void *p),
2565                              void *priv)
2566 {
2567         struct threads *threads;
2568         struct rb_node *nd;
2569         struct thread *thread;
2570         int rc = 0;
2571         int i;
2572
2573         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2574                 threads = &machine->threads[i];
2575                 for (nd = rb_first_cached(&threads->entries); nd;
2576                      nd = rb_next(nd)) {
2577                         thread = rb_entry(nd, struct thread, rb_node);
2578                         rc = fn(thread, priv);
2579                         if (rc != 0)
2580                                 return rc;
2581                 }
2582
2583                 list_for_each_entry(thread, &threads->dead, node) {
2584                         rc = fn(thread, priv);
2585                         if (rc != 0)
2586                                 return rc;
2587                 }
2588         }
2589         return rc;
2590 }
2591
2592 int machines__for_each_thread(struct machines *machines,
2593                               int (*fn)(struct thread *thread, void *p),
2594                               void *priv)
2595 {
2596         struct rb_node *nd;
2597         int rc = 0;
2598
2599         rc = machine__for_each_thread(&machines->host, fn, priv);
2600         if (rc != 0)
2601                 return rc;
2602
2603         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2604                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2605
2606                 rc = machine__for_each_thread(machine, fn, priv);
2607                 if (rc != 0)
2608                         return rc;
2609         }
2610         return rc;
2611 }
2612
2613 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2614 {
2615         int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2616
2617         if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
2618                 return -1;
2619
2620         return machine->current_tid[cpu];
2621 }
2622
2623 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2624                              pid_t tid)
2625 {
2626         struct thread *thread;
2627         int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2628
2629         if (cpu < 0)
2630                 return -EINVAL;
2631
2632         if (!machine->current_tid) {
2633                 int i;
2634
2635                 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
2636                 if (!machine->current_tid)
2637                         return -ENOMEM;
2638                 for (i = 0; i < nr_cpus; i++)
2639                         machine->current_tid[i] = -1;
2640         }
2641
2642         if (cpu >= nr_cpus) {
2643                 pr_err("Requested CPU %d too large. ", cpu);
2644                 pr_err("Consider raising MAX_NR_CPUS\n");
2645                 return -EINVAL;
2646         }
2647
2648         machine->current_tid[cpu] = tid;
2649
2650         thread = machine__findnew_thread(machine, pid, tid);
2651         if (!thread)
2652                 return -ENOMEM;
2653
2654         thread->cpu = cpu;
2655         thread__put(thread);
2656
2657         return 0;
2658 }
2659
2660 /*
2661  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2662  * normalized arch is needed.
2663  */
2664 bool machine__is(struct machine *machine, const char *arch)
2665 {
2666         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2667 }
2668
2669 int machine__nr_cpus_avail(struct machine *machine)
2670 {
2671         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2672 }
2673
2674 int machine__get_kernel_start(struct machine *machine)
2675 {
2676         struct map *map = machine__kernel_map(machine);
2677         int err = 0;
2678
2679         /*
2680          * The only addresses above 2^63 are kernel addresses of a 64-bit
2681          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2682          * all addresses including kernel addresses are less than 2^32.  In
2683          * that case (32-bit system), if the kernel mapping is unknown, all
2684          * addresses will be assumed to be in user space - see
2685          * machine__kernel_ip().
2686          */
2687         machine->kernel_start = 1ULL << 63;
2688         if (map) {
2689                 err = map__load(map);
2690                 /*
2691                  * On x86_64, PTI entry trampolines are less than the
2692                  * start of kernel text, but still above 2^63. So leave
2693                  * kernel_start = 1ULL << 63 for x86_64.
2694                  */
2695                 if (!err && !machine__is(machine, "x86_64"))
2696                         machine->kernel_start = map->start;
2697         }
2698         return err;
2699 }
2700
2701 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2702 {
2703         u8 addr_cpumode = cpumode;
2704         bool kernel_ip;
2705
2706         if (!machine->single_address_space)
2707                 goto out;
2708
2709         kernel_ip = machine__kernel_ip(machine, addr);
2710         switch (cpumode) {
2711         case PERF_RECORD_MISC_KERNEL:
2712         case PERF_RECORD_MISC_USER:
2713                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2714                                            PERF_RECORD_MISC_USER;
2715                 break;
2716         case PERF_RECORD_MISC_GUEST_KERNEL:
2717         case PERF_RECORD_MISC_GUEST_USER:
2718                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2719                                            PERF_RECORD_MISC_GUEST_USER;
2720                 break;
2721         default:
2722                 break;
2723         }
2724 out:
2725         return addr_cpumode;
2726 }
2727
2728 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2729 {
2730         return dsos__findnew(&machine->dsos, filename);
2731 }
2732
2733 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2734 {
2735         struct machine *machine = vmachine;
2736         struct map *map;
2737         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2738
2739         if (sym == NULL)
2740                 return NULL;
2741
2742         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2743         *addrp = map->unmap_ip(map, sym->start);
2744         return sym->name;
2745 }